📊 Context Engineering • Tool Definition Tax • Anthropic Prompt Caching • Lost-in-the-Middle
AI Agent Context & Token Budget Optimizer
Agent tool schemas and multi-turn message history burn context tokens fast. Allocate your token budget across system instructions, tool definitions, RAG context, and safety buffers with automated cache break-even forecasting.
⚙️ Token Allocation Inputs 128,000 max
1. System Instructions & Rules 2,500 tok
2. MCP / Tool Definitions JSON Tax 3,200 tok
Approx. ~350 tokens per typical MCP tool function schema
3. RAG Document & Vector Chunks 12,000 tok
4. Multi-Turn Message History 6,000 tok
5. Reserved Max Generation Tokens 4,096 tok
Context Window Distribution
Using 27,796 tokens of 200,000 (13.9% capacity)
🟢 Optimal Attention Recall
System Prompt (2.5k)
Tool Schemas (3.2k)
RAG Retrieval (12.0k)
Chat History (6.0k)
Reserved Output (4.1k)
⚡ Prompt Caching Impact Supported (Anthropic 90% Read Discount)
Cold Turn Cost
$0.083
Zero cache hits
Cached Turn Cost
$0.012
85% discount on prefix
1,000 Turn Savings
+$71.40
Monthly cost reduction